#include "cfgloop.h"
#include "expr.h"
#include "optabs.h"
+#include "recog.h"
#include "tree-data-ref.h"
#include "tree-chrec.h"
#include "tree-scalar-evolution.h"
static tree vect_create_destination_var (tree, tree);
static tree vect_create_data_ref_ptr
(tree, block_stmt_iterator *, tree, tree *, bool);
-static tree vect_create_index_for_vector_ref (loop_vec_info);
static tree vect_create_addr_base_for_vector_ref (tree, tree *, tree);
static tree vect_get_new_vect_var (tree, enum vect_var_kind, const char *);
static tree vect_get_vec_def_for_operand (tree, tree, tree *);
}
-/* Function vect_create_index_for_vector_ref.
-
- Create (and return) an index variable, along with it's update chain in the
- loop. This variable will be used to access a memory location in a vector
- operation.
-
- Input:
- LOOP: The loop being vectorized.
- BSI: The block_stmt_iterator where STMT is. Any new stmts created by this
- function can be added here, or in the loop pre-header.
-
- Output:
- Return an index that will be used to index a vector array. It is expected
- that a pointer to the first vector will be used as the base address for the
- indexed reference.
-
- FORNOW: we are not trying to be efficient, just creating a new index each
- time from scratch. At this time all vector references could use the same
- index.
-
- TODO: create only one index to be used by all vector references. Record
- the index in the LOOP_VINFO the first time this procedure is called and
- return it on subsequent calls. The increment of this index must be placed
- just before the conditional expression that ends the single block loop. */
-
-static tree
-vect_create_index_for_vector_ref (loop_vec_info loop_vinfo)
-{
- tree init, step;
- block_stmt_iterator incr_bsi;
- bool insert_after;
- tree indx_before_incr, indx_after_incr;
- struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
- tree incr;
-
- /* It is assumed that the base pointer used for vectorized access contains
- the address of the first vector. Therefore the index used for vectorized
- access must be initialized to zero and incremented by 1. */
-
- init = integer_zero_node;
- step = integer_one_node;
-
- standard_iv_increment_position (loop, &incr_bsi, &insert_after);
- create_iv (init, step, NULL_TREE, loop, &incr_bsi, insert_after,
- &indx_before_incr, &indx_after_incr);
- incr = bsi_stmt (incr_bsi);
- set_stmt_info ((tree_ann_t)stmt_ann (incr), new_stmt_vec_info (incr, loop_vinfo));
-
- return indx_before_incr;
-}
-
-
/* Function vect_create_addr_base_for_vector_ref.
Create an expression that computes the address of the first memory location
{
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
- tree data_ref_base =
- unshare_expr (STMT_VINFO_VECT_DR_BASE_ADDRESS (stmt_info));
+ tree data_ref_base = unshare_expr (DR_BASE_ADDRESS (dr));
tree base_name = build_fold_indirect_ref (data_ref_base);
tree ref = DR_REF (dr);
tree scalar_type = TREE_TYPE (ref);
tree new_temp;
tree addr_base, addr_expr;
tree dest, new_stmt;
- tree base_offset = unshare_expr (STMT_VINFO_VECT_INIT_OFFSET (stmt_info));
+ tree base_offset = unshare_expr (DR_OFFSET (dr));
+ tree init = unshare_expr (DR_INIT (dr));
/* Create base_offset */
+ base_offset = size_binop (PLUS_EXPR, base_offset, init);
dest = create_tmp_var (TREE_TYPE (base_offset), "base_off");
add_referenced_tmp_var (dest);
base_offset = force_gimple_operand (base_offset, &new_stmt, false, dest);
tree tmp = create_tmp_var (TREE_TYPE (base_offset), "offset");
add_referenced_tmp_var (tmp);
offset = fold_build2 (MULT_EXPR, TREE_TYPE (offset), offset,
- STMT_VINFO_VECT_STEP (stmt_info));
+ DR_STEP (dr));
base_offset = fold_build2 (PLUS_EXPR, TREE_TYPE (base_offset),
base_offset, offset);
base_offset = force_gimple_operand (base_offset, &new_stmt, false, tmp);
Return the initial_address in INITIAL_ADDRESS.
- 2. Create a data-reference in the loop based on the new vector pointer vp,
- and using a new index variable 'idx' as follows:
-
- vp' = vp + update
-
- where if ONLY_INIT is true:
- update = zero
- and otherwise
- update = idx + vector_type_size
-
- Return the pointer vp'.
-
+ 2. If ONLY_INIT is true, return the initial pointer. Otherwise, create
+ a data-reference in the loop based on the new vector pointer vp. This
+ new data reference will by some means be updated each iteration of
+ the loop. Return the pointer vp'.
FORNOW: handle only aligned and consecutive accesses. */
static tree
-vect_create_data_ref_ptr (tree stmt, block_stmt_iterator *bsi, tree offset,
- tree *initial_address, bool only_init)
+vect_create_data_ref_ptr (tree stmt,
+ block_stmt_iterator *bsi ATTRIBUTE_UNUSED,
+ tree offset, tree *initial_address, bool only_init)
{
tree base_name;
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
tree new_temp;
tree vec_stmt;
tree new_stmt_list = NULL_TREE;
- tree idx;
edge pe = loop_preheader_edge (loop);
basic_block new_bb;
tree vect_ptr_init;
- tree vectype_size;
- tree ptr_update;
- tree data_ref_ptr;
- tree type, tmp, size;
+ struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
- base_name = build_fold_indirect_ref (unshare_expr (
- STMT_VINFO_VECT_DR_BASE_ADDRESS (stmt_info)));
+ base_name = build_fold_indirect_ref (unshare_expr (DR_BASE_ADDRESS (dr)));
if (vect_print_dump_info (REPORT_DETAILS))
{
tree data_ref_base = base_name;
- fprintf (vect_dump, "create array_ref of type: ");
+ fprintf (vect_dump, "create vector-pointer variable to type: ");
print_generic_expr (vect_dump, vectype, TDF_SLIM);
if (TREE_CODE (data_ref_base) == VAR_DECL)
fprintf (vect_dump, " vectorizing a one dimensional array ref: ");
/** (2) Add aliasing information to the new vector-pointer:
- (The points-to info (SSA_NAME_PTR_INFO) may be defined later.) **/
+ (The points-to info (DR_PTR_INFO) may be defined later.) **/
- tag = STMT_VINFO_MEMTAG (stmt_info);
+ tag = DR_MEMTAG (dr);
gcc_assert (tag);
/* If tag is a variable (and NOT_A_TAG) than a new type alias
else
var_ann (vect_ptr)->type_mem_tag = tag;
- var_ann (vect_ptr)->subvars = STMT_VINFO_SUBVARS (stmt_info);
+ var_ann (vect_ptr)->subvars = DR_SUBVARS (dr);
/** (3) Calculate the initial address the vector-pointer, and set
the vector-pointer to point to it before the loop: **/
/* Create: p = (vectype *) initial_base */
vec_stmt = fold_convert (vect_ptr_type, new_temp);
vec_stmt = build2 (MODIFY_EXPR, void_type_node, vect_ptr, vec_stmt);
- new_temp = make_ssa_name (vect_ptr, vec_stmt);
- TREE_OPERAND (vec_stmt, 0) = new_temp;
+ vect_ptr_init = make_ssa_name (vect_ptr, vec_stmt);
+ TREE_OPERAND (vec_stmt, 0) = vect_ptr_init;
new_bb = bsi_insert_on_edge_immediate (pe, vec_stmt);
gcc_assert (!new_bb);
- vect_ptr_init = TREE_OPERAND (vec_stmt, 0);
/** (4) Handle the updating of the vector-pointer inside the loop: **/
if (only_init) /* No update in loop is required. */
{
/* Copy the points-to information if it exists. */
- if (STMT_VINFO_PTR_INFO (stmt_info))
- duplicate_ssa_name_ptr_info (vect_ptr_init,
- STMT_VINFO_PTR_INFO (stmt_info));
+ if (DR_PTR_INFO (dr))
+ duplicate_ssa_name_ptr_info (vect_ptr_init, DR_PTR_INFO (dr));
return vect_ptr_init;
}
+ else
+ {
+ block_stmt_iterator incr_bsi;
+ bool insert_after;
+ tree indx_before_incr, indx_after_incr;
+ tree incr;
+
+ standard_iv_increment_position (loop, &incr_bsi, &insert_after);
+ create_iv (vect_ptr_init,
+ fold_convert (vect_ptr_type, TYPE_SIZE_UNIT (vectype)),
+ NULL_TREE, loop, &incr_bsi, insert_after,
+ &indx_before_incr, &indx_after_incr);
+ incr = bsi_stmt (incr_bsi);
+ set_stmt_info ((tree_ann_t)stmt_ann (incr),
+ new_stmt_vec_info (incr, loop_vinfo));
- idx = vect_create_index_for_vector_ref (loop_vinfo);
-
- /* Create: update = idx * vectype_size */
- tmp = create_tmp_var (integer_type_node, "update");
- add_referenced_tmp_var (tmp);
- size = TYPE_SIZE (vect_ptr_type);
- type = lang_hooks.types.type_for_size (tree_low_cst (size, 1), 1);
- ptr_update = create_tmp_var (type, "update");
- add_referenced_tmp_var (ptr_update);
- vectype_size = TYPE_SIZE_UNIT (vectype);
- vec_stmt = build2 (MULT_EXPR, integer_type_node, idx, vectype_size);
- vec_stmt = build2 (MODIFY_EXPR, void_type_node, tmp, vec_stmt);
- new_temp = make_ssa_name (tmp, vec_stmt);
- TREE_OPERAND (vec_stmt, 0) = new_temp;
- bsi_insert_before (bsi, vec_stmt, BSI_SAME_STMT);
- vec_stmt = fold_convert (type, new_temp);
- vec_stmt = build2 (MODIFY_EXPR, void_type_node, ptr_update, vec_stmt);
- new_temp = make_ssa_name (ptr_update, vec_stmt);
- TREE_OPERAND (vec_stmt, 0) = new_temp;
- bsi_insert_before (bsi, vec_stmt, BSI_SAME_STMT);
-
- /* Create: data_ref_ptr = vect_ptr_init + update */
- vec_stmt = build2 (PLUS_EXPR, vect_ptr_type, vect_ptr_init, new_temp);
- vec_stmt = build2 (MODIFY_EXPR, void_type_node, vect_ptr, vec_stmt);
- new_temp = make_ssa_name (vect_ptr, vec_stmt);
- TREE_OPERAND (vec_stmt, 0) = new_temp;
- bsi_insert_before (bsi, vec_stmt, BSI_SAME_STMT);
- data_ref_ptr = TREE_OPERAND (vec_stmt, 0);
+ /* Copy the points-to information if it exists. */
+ if (DR_PTR_INFO (dr))
+ {
+ duplicate_ssa_name_ptr_info (indx_before_incr, DR_PTR_INFO (dr));
+ duplicate_ssa_name_ptr_info (indx_after_incr, DR_PTR_INFO (dr));
+ }
+ merge_alias_info (vect_ptr_init, indx_before_incr);
+ merge_alias_info (vect_ptr_init, indx_after_incr);
- /* Copy the points-to information if it exists. */
- if (STMT_VINFO_PTR_INFO (stmt_info))
- duplicate_ssa_name_ptr_info (data_ref_ptr, STMT_VINFO_PTR_INFO (stmt_info));
- return data_ref_ptr;
+ return indx_before_incr;
+ }
}
t = tree_cons (NULL_TREE, def, t);
}
- vec_inv = build_constructor (vectype, t);
+ /* FIXME: use build_constructor directly. */
+ vec_inv = build_constructor_from_list (vectype, t);
return vect_init_vector (stmt, vec_inv);
}
case MAX_EXPR:
def = init_val;
nelements = nunits;
- need_epilog_adjust = true;
+ need_epilog_adjust = false;
break;
default:
if (TREE_CODE (init_val) == INTEGER_CST || TREE_CODE (init_val) == REAL_CST)
vec = build_vector (vectype, t);
else
- vec = build_constructor (vectype, t);
+ vec = build_constructor_from_list (vectype, t);
if (!need_epilog_adjust)
- {
- if (INTEGRAL_TYPE_P (type))
- init_val = build_int_cst (type, 0);
- else
- init_val = build_real (type, dconst0);
- }
- *scalar_def = init_val;
+ *scalar_def = NULL_TREE;
+ else
+ *scalar_def = init_val;
return vect_init_vector (stmt, vec);
}
imm_use_iterator imm_iter;
use_operand_p use_p;
bool extract_scalar_result;
- bool adjust_in_epilog;
/*** 1. Create the reduction def-use cycle ***/
bsi_insert_after (&exit_bsi, epilog_stmt, BSI_NEW_STMT);
extract_scalar_result = true;
- adjust_in_epilog = true;
}
else
{
- enum tree_code shift_code;
+ enum tree_code shift_code = 0;
bool have_whole_vector_shift = true;
enum tree_code code = TREE_CODE (TREE_OPERAND (stmt, 1)); /* CHECKME */
int bit_offset;
}
extract_scalar_result = true;
- adjust_in_epilog = true;
}
else
{
+ tree rhs;
+
/*** Case 3:
- Create: s = init;
- for (offset=0; offset<vector_size; offset+=element_size;)
+ Create:
+ s = extract_field <v_out2, 0>
+ for (offset=element_size; offset<vector_size; offset+=element_size;)
{
Create: s' = extract_field <v_out2, offset>
Create: s = op <s, s'>
vec_temp = PHI_RESULT (new_phi);
vec_size_in_bits = tree_low_cst (TYPE_SIZE (vectype), 1);
- /* first iteration is peeled out when possible to minimize
- the number of operations we generate: */
- if (code == PLUS_EXPR
- && (integer_zerop (scalar_initial_def)
- || real_zerop (scalar_initial_def)))
- {
- epilog_stmt = build2 (MODIFY_EXPR, scalar_type, new_scalar_dest,
- build3 (BIT_FIELD_REF, scalar_type,
- vec_temp, bitsize, bitsize_zero_node));
- new_temp = make_ssa_name (new_scalar_dest, epilog_stmt);
- TREE_OPERAND (epilog_stmt, 0) = new_temp;
- bsi_insert_after (&exit_bsi, epilog_stmt, BSI_NEW_STMT);
- if (vect_print_dump_info (REPORT_DETAILS))
- print_generic_expr (vect_dump, epilog_stmt, TDF_SLIM);
-
- bit_offset = element_bitsize;
- }
- else
- {
- new_temp = scalar_initial_def;
- bit_offset = 0;
- }
+ rhs = build3 (BIT_FIELD_REF, scalar_type, vec_temp, bitsize,
+ bitsize_zero_node);
- for (;
+ BIT_FIELD_REF_UNSIGNED (rhs) = TYPE_UNSIGNED (scalar_type);
+ epilog_stmt = build2 (MODIFY_EXPR, scalar_type, new_scalar_dest,
+ rhs);
+ new_temp = make_ssa_name (new_scalar_dest, epilog_stmt);
+ TREE_OPERAND (epilog_stmt, 0) = new_temp;
+ bsi_insert_after (&exit_bsi, epilog_stmt, BSI_NEW_STMT);
+ if (vect_print_dump_info (REPORT_DETAILS))
+ print_generic_expr (vect_dump, epilog_stmt, TDF_SLIM);
+
+ for (bit_offset = element_bitsize;
bit_offset < vec_size_in_bits;
bit_offset += element_bitsize)
{
tree bitpos = bitsize_int (bit_offset);
-
+ tree rhs = build3 (BIT_FIELD_REF, scalar_type, vec_temp, bitsize,
+ bitpos);
+
+ BIT_FIELD_REF_UNSIGNED (rhs) = TYPE_UNSIGNED (scalar_type);
epilog_stmt = build2 (MODIFY_EXPR, scalar_type, new_scalar_dest,
- build3 (BIT_FIELD_REF, scalar_type,
- vec_temp, bitsize, bitpos));
+ rhs);
new_name = make_ssa_name (new_scalar_dest, epilog_stmt);
TREE_OPERAND (epilog_stmt, 0) = new_name;
bsi_insert_after (&exit_bsi, epilog_stmt, BSI_NEW_STMT);
}
extract_scalar_result = false;
- adjust_in_epilog = false;
}
}
if (extract_scalar_result)
{
+ tree rhs;
+
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "extract scalar result");
/* The result is in the low order bits. */
- if (BITS_BIG_ENDIAN)
+ if (BYTES_BIG_ENDIAN)
bitpos = size_binop (MULT_EXPR,
bitsize_int (TYPE_VECTOR_SUBPARTS (vectype) - 1),
TYPE_SIZE (scalar_type));
else
bitpos = bitsize_zero_node;
- epilog_stmt = build2 (MODIFY_EXPR, scalar_type, new_scalar_dest,
- build3 (BIT_FIELD_REF, scalar_type,
- new_temp, bitsize, bitpos));
+ rhs = build3 (BIT_FIELD_REF, scalar_type, new_temp, bitsize, bitpos);
+ BIT_FIELD_REF_UNSIGNED (rhs) = TYPE_UNSIGNED (scalar_type);
+ epilog_stmt = build2 (MODIFY_EXPR, scalar_type, new_scalar_dest, rhs);
new_temp = make_ssa_name (new_scalar_dest, epilog_stmt);
TREE_OPERAND (epilog_stmt, 0) = new_temp;
bsi_insert_after (&exit_bsi, epilog_stmt, BSI_NEW_STMT);
Create:
s_out = scalar_expr <s_out, scalar_initial_def> */
- if (adjust_in_epilog)
+ if (scalar_initial_def)
{
epilog_stmt = build2 (MODIFY_EXPR, scalar_type, new_scalar_dest,
build2 (code, scalar_type, new_temp, scalar_initial_def));
/* Prepare the operand that is defined inside the loop body */
loop_vec_def = vect_get_vec_def_for_operand (op0, stmt, NULL);
- gcc_assert (VECTOR_MODE_P (TYPE_MODE (TREE_TYPE (loop_vec_def))));
-
/* Create the vectorized operation that computes the partial results */
*vec_stmt = build2 (MODIFY_EXPR, vectype, vec_dest,
int op_type;
tree op;
optab optab;
+ int icode;
+ enum machine_mode optab_op2_mode;
tree def, def_stmt;
enum vect_def_type dt;
return false;
}
vec_mode = TYPE_MODE (vectype);
- if (optab->handlers[(int) vec_mode].insn_code == CODE_FOR_nothing)
+ icode = (int) optab->handlers[(int) vec_mode].insn_code;
+ if (icode == CODE_FOR_nothing)
{
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "op not supported by target.");
return false;
}
+ if (code == LSHIFT_EXPR || code == RSHIFT_EXPR)
+ {
+ /* FORNOW: not yet supported. */
+ if (!VECTOR_MODE_P (vec_mode))
+ return false;
+
+ /* Invariant argument is needed for a vector shift
+ by a scalar shift operand. */
+ optab_op2_mode = insn_data[icode].operand[2].mode;
+ if (! (VECTOR_MODE_P (optab_op2_mode)
+ || dt == vect_constant_def
+ || dt == vect_invariant_def))
+ {
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "operand mode requires invariant argument.");
+ return false;
+ }
+ }
+
if (!vec_stmt) /* transformation not required. */
{
STMT_VINFO_TYPE (stmt_info) = op_vec_info_type;
if (op_type == binary_op)
{
op1 = TREE_OPERAND (operation, 1);
- vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, NULL);
+
+ if (code == LSHIFT_EXPR || code == RSHIFT_EXPR)
+ {
+ /* Vector shl and shr insn patterns can be defined with
+ scalar operand 2 (shift operand). In this case, use
+ constant or loop invariant op1 directly, without
+ extending it to vector mode first. */
+
+ optab_op2_mode = insn_data[icode].operand[2].mode;
+ if (!VECTOR_MODE_P (optab_op2_mode))
+ {
+ if (vect_print_dump_info (REPORT_DETAILS))
+ fprintf (vect_dump, "operand 1 using scalar mode.");
+ vec_oprnd1 = op1;
+ }
+ }
+
+ if (!vec_oprnd1)
+ vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, NULL);
}
/* Arguments are ready. create the new vector stmt. */
/* <2> Create lsq = *(floor(p2')) in the loop */
- offset = build_int_cst (integer_type_node,
- TYPE_VECTOR_SUBPARTS (vectype));
- offset = int_const_binop (MINUS_EXPR, offset, integer_one_node, 1);
+ offset = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1);
vec_dest = vect_create_destination_var (scalar_dest, vectype);
dataref_ptr = vect_create_data_ref_ptr (stmt, bsi, offset, &dummy, false);
data_ref = build1 (ALIGN_INDIRECT_REF, vectype, dataref_ptr);
the value of the parameter and no global variables are touched
which makes the builtin a "const" function. Requiring the
builtin to have the "const" attribute makes it unnecessary
- to call mark_call_clobbered_vars_to_rename. */
+ to call mark_call_clobbered. */
gcc_assert (TREE_READONLY (builtin_decl));
}
else
/* Arguments are ready. create the new vector stmt. */
vec_compare = build2 (TREE_CODE (cond_expr), vectype,
vec_cond_lhs, vec_cond_rhs);
- vec_cond_expr = build (VEC_COND_EXPR, vectype,
- vec_compare, vec_then_clause, vec_else_clause);
+ vec_cond_expr = build3 (VEC_COND_EXPR, vectype,
+ vec_compare, vec_then_clause, vec_else_clause);
*vec_stmt = build2 (MODIFY_EXPR, vectype, vec_dest, vec_cond_expr);
new_temp = make_ssa_name (vec_dest, *vec_stmt);
struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
tree ni = LOOP_VINFO_NITERS (loop_vinfo);
int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
- tree log_vf = build_int_cst (unsigned_type_node, exact_log2 (vf));
+ tree log_vf;
pe = loop_preheader_edge (loop);
number of iterations loop executes. */
ni_name = vect_build_loop_niters (loop_vinfo);
+ log_vf = build_int_cst (TREE_TYPE (ni), exact_log2 (vf));
/* Create: ratio = ni >> log2(vf) */
stmt_vec_info stmt_info = vinfo_for_stmt (dr_stmt);
tree vectype = STMT_VINFO_VECTYPE (stmt_info);
int vectype_align = TYPE_ALIGN (vectype) / BITS_PER_UNIT;
- tree vf_minus_1 = build_int_cst (unsigned_type_node, vf - 1);
tree niters_type = TREE_TYPE (loop_niters);
pe = loop_preheader_edge (loop);
tree type = lang_hooks.types.type_for_size (tree_low_cst (size, 1), 1);
tree vectype_size_minus_1 = build_int_cst (type, vectype_align - 1);
tree elem_size_log =
- build_int_cst (unsigned_type_node, exact_log2 (vectype_align/vf));
- tree vf_tree = build_int_cst (unsigned_type_node, vf);
+ build_int_cst (type, exact_log2 (vectype_align/vf));
+ tree vf_minus_1 = build_int_cst (type, vf - 1);
+ tree vf_tree = build_int_cst (type, vf);
tree byte_misalign;
tree elem_misalign;
/* Create: elem_misalign = byte_misalign / element_size */
elem_misalign =
- build2 (RSHIFT_EXPR, unsigned_type_node, byte_misalign, elem_size_log);
+ build2 (RSHIFT_EXPR, type, byte_misalign, elem_size_log);
/* Create: (niters_type) (VF - elem_misalign)&(VF - 1) */
- iters = build2 (MINUS_EXPR, unsigned_type_node, vf_tree, elem_misalign);
- iters = build2 (BIT_AND_EXPR, unsigned_type_node, iters, vf_minus_1);
+ iters = build2 (MINUS_EXPR, type, vf_tree, elem_misalign);
+ iters = build2 (BIT_AND_EXPR, type, iters, vf_minus_1);
iters = fold_convert (niters_type, iters);
}
NITERS iterations were peeled from LOOP. DR represents a data reference
in LOOP. This function updates the information recorded in DR to
account for the fact that the first NITERS iterations had already been
- executed. Specifically, it updates the OFFSET field of stmt_info. */
+ executed. Specifically, it updates the OFFSET field of DR. */
static void
vect_update_init_of_dr (struct data_reference *dr, tree niters)
{
- stmt_vec_info stmt_info = vinfo_for_stmt (DR_STMT (dr));
- tree offset = STMT_VINFO_VECT_INIT_OFFSET (stmt_info);
+ tree offset = DR_OFFSET (dr);
- niters = fold_build2 (MULT_EXPR, TREE_TYPE (niters), niters,
- STMT_VINFO_VECT_STEP (stmt_info));
+ niters = fold_build2 (MULT_EXPR, TREE_TYPE (niters), niters, DR_STEP (dr));
offset = fold_build2 (PLUS_EXPR, TREE_TYPE (offset), offset, niters);
- STMT_VINFO_VECT_INIT_OFFSET (stmt_info) = offset;
+ DR_OFFSET (dr) = offset;
}
vect_update_inits_of_drs (loop_vec_info loop_vinfo, tree niters)
{
unsigned int i;
- varray_type loop_write_datarefs = LOOP_VINFO_DATAREF_WRITES (loop_vinfo);
- varray_type loop_read_datarefs = LOOP_VINFO_DATAREF_READS (loop_vinfo);
+ varray_type datarefs = LOOP_VINFO_DATAREFS (loop_vinfo);
if (vect_dump && (dump_flags & TDF_DETAILS))
fprintf (vect_dump, "=== vect_update_inits_of_dr ===");
- for (i = 0; i < VARRAY_ACTIVE_SIZE (loop_write_datarefs); i++)
- {
- struct data_reference *dr = VARRAY_GENERIC_PTR (loop_write_datarefs, i);
- vect_update_init_of_dr (dr, niters);
- }
-
- for (i = 0; i < VARRAY_ACTIVE_SIZE (loop_read_datarefs); i++)
+ for (i = 0; i < VARRAY_ACTIVE_SIZE (datarefs); i++)
{
- struct data_reference *dr = VARRAY_GENERIC_PTR (loop_read_datarefs, i);
+ struct data_reference *dr = VARRAY_GENERIC_PTR (datarefs, i);
vect_update_init_of_dr (dr, niters);
}
}
}
+/* Function vect_create_cond_for_align_checks.
+
+ Create a conditional expression that represents the alignment checks for
+ all of data references (array element references) whose alignment must be
+ checked at runtime.
+
+ Input:
+ LOOP_VINFO - two fields of the loop information are used.
+ LOOP_VINFO_PTR_MASK is the mask used to check the alignment.
+ LOOP_VINFO_MAY_MISALIGN_STMTS contains the refs to be checked.
+
+ Output:
+ COND_EXPR_STMT_LIST - statements needed to construct the conditional
+ expression.
+ The returned value is the conditional expression to be used in the if
+ statement that controls which version of the loop gets executed at runtime.
+
+ The algorithm makes two assumptions:
+ 1) The number of bytes "n" in a vector is a power of 2.
+ 2) An address "a" is aligned if a%n is zero and that this
+ test can be done as a&(n-1) == 0. For example, for 16
+ byte vectors the test is a&0xf == 0. */
+
+static tree
+vect_create_cond_for_align_checks (loop_vec_info loop_vinfo,
+ tree *cond_expr_stmt_list)
+{
+ VEC(tree,heap) *may_misalign_stmts
+ = LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo);
+ tree ref_stmt;
+ int mask = LOOP_VINFO_PTR_MASK (loop_vinfo);
+ tree mask_cst;
+ unsigned int i;
+ tree psize;
+ tree int_ptrsize_type;
+ char tmp_name[20];
+ tree or_tmp_name = NULL_TREE;
+ tree and_tmp, and_tmp_name, and_stmt;
+ tree ptrsize_zero;
+
+ /* Check that mask is one less than a power of 2, i.e., mask is
+ all zeros followed by all ones. */
+ gcc_assert ((mask != 0) && ((mask & (mask+1)) == 0));
+
+ /* CHECKME: what is the best integer or unsigned type to use to hold a
+ cast from a pointer value? */
+ psize = TYPE_SIZE (ptr_type_node);
+ int_ptrsize_type
+ = lang_hooks.types.type_for_size (tree_low_cst (psize, 1), 0);
+
+ /* Create expression (mask & (dr_1 || ... || dr_n)) where dr_i is the address
+ of the first vector of the i'th data reference. */
+
+ for (i = 0; VEC_iterate (tree, may_misalign_stmts, i, ref_stmt); i++)
+ {
+ tree new_stmt_list = NULL_TREE;
+ tree addr_base;
+ tree addr_tmp, addr_tmp_name, addr_stmt;
+ tree or_tmp, new_or_tmp_name, or_stmt;
+
+ /* create: addr_tmp = (int)(address_of_first_vector) */
+ addr_base = vect_create_addr_base_for_vector_ref (ref_stmt,
+ &new_stmt_list,
+ NULL_TREE);
+
+ if (new_stmt_list != NULL_TREE)
+ append_to_statement_list_force (new_stmt_list, cond_expr_stmt_list);
+
+ sprintf (tmp_name, "%s%d", "addr2int", i);
+ addr_tmp = create_tmp_var (int_ptrsize_type, tmp_name);
+ add_referenced_tmp_var (addr_tmp);
+ addr_tmp_name = make_ssa_name (addr_tmp, NULL_TREE);
+ addr_stmt = fold_convert (int_ptrsize_type, addr_base);
+ addr_stmt = build2 (MODIFY_EXPR, void_type_node,
+ addr_tmp_name, addr_stmt);
+ SSA_NAME_DEF_STMT (addr_tmp_name) = addr_stmt;
+ append_to_statement_list_force (addr_stmt, cond_expr_stmt_list);
+
+ /* The addresses are OR together. */
+
+ if (or_tmp_name != NULL_TREE)
+ {
+ /* create: or_tmp = or_tmp | addr_tmp */
+ sprintf (tmp_name, "%s%d", "orptrs", i);
+ or_tmp = create_tmp_var (int_ptrsize_type, tmp_name);
+ add_referenced_tmp_var (or_tmp);
+ new_or_tmp_name = make_ssa_name (or_tmp, NULL_TREE);
+ or_stmt = build2 (MODIFY_EXPR, void_type_node, new_or_tmp_name,
+ build2 (BIT_IOR_EXPR, int_ptrsize_type,
+ or_tmp_name,
+ addr_tmp_name));
+ SSA_NAME_DEF_STMT (new_or_tmp_name) = or_stmt;
+ append_to_statement_list_force (or_stmt, cond_expr_stmt_list);
+ or_tmp_name = new_or_tmp_name;
+ }
+ else
+ or_tmp_name = addr_tmp_name;
+
+ } /* end for i */
+
+ mask_cst = build_int_cst (int_ptrsize_type, mask);
+
+ /* create: and_tmp = or_tmp & mask */
+ and_tmp = create_tmp_var (int_ptrsize_type, "andmask" );
+ add_referenced_tmp_var (and_tmp);
+ and_tmp_name = make_ssa_name (and_tmp, NULL_TREE);
+
+ and_stmt = build2 (MODIFY_EXPR, void_type_node,
+ and_tmp_name,
+ build2 (BIT_AND_EXPR, int_ptrsize_type,
+ or_tmp_name, mask_cst));
+ SSA_NAME_DEF_STMT (and_tmp_name) = and_stmt;
+ append_to_statement_list_force (and_stmt, cond_expr_stmt_list);
+
+ /* Make and_tmp the left operand of the conditional test against zero.
+ if and_tmp has a non-zero bit then some address is unaligned. */
+ ptrsize_zero = build_int_cst (int_ptrsize_type, 0);
+ return build2 (EQ_EXPR, boolean_type_node,
+ and_tmp_name, ptrsize_zero);
+}
+
+
/* Function vect_transform_loop.
The analysis phase has determined that the loop is vectorizable.
int i;
tree ratio = NULL;
int vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
+ bitmap_iterator bi;
+ unsigned int j;
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "=== vec_transform_loop ===");
+ /* If the loop has data references that may or may not be aligned then
+ two versions of the loop need to be generated, one which is vectorized
+ and one which isn't. A test is then generated to control which of the
+ loops is executed. The test checks for the alignment of all of the
+ data references that may or may not be aligned. */
+
+ if (VEC_length (tree, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo)))
+ {
+ struct loop *nloop;
+ tree cond_expr;
+ tree cond_expr_stmt_list = NULL_TREE;
+ basic_block condition_bb;
+ block_stmt_iterator cond_exp_bsi;
+
+ cond_expr = vect_create_cond_for_align_checks (loop_vinfo,
+ &cond_expr_stmt_list);
+ initialize_original_copy_tables ();
+ nloop = loop_version (loops, loop, cond_expr, &condition_bb, true);
+ free_original_copy_tables();
+ update_ssa (TODO_update_ssa);
+ cond_exp_bsi = bsi_last (condition_bb);
+ bsi_insert_before (&cond_exp_bsi, cond_expr_stmt_list, BSI_SAME_STMT);
+ }
+
+ /* CHECKME: we wouldn't need this if we calles update_ssa once
+ for all loops. */
+ bitmap_zero (vect_vnames_to_rename);
+
/* Peel the loop if there are data refs with unknown alignment.
Only one data ref with unknown store is allowed. */
slpeel_make_loop_iterate_ntimes (loop, ratio);
+ EXECUTE_IF_SET_IN_BITMAP (vect_vnames_to_rename, 0, j, bi)
+ mark_sym_for_renaming (SSA_NAME_VAR (ssa_name (j)));
+
/* The memory tags and pointers in vectorized statements need to
have their SSA forms updated. FIXME, why can't this be delayed
until all the loops have been transformed? */